Remote thin clients are the best forward for most usecases and enthusiasts and goymers can't stand it for some reason. Why are we wasting so many resources making desktops and laptops and consoles that squeeze maybe 5% of the potential of the silicon over

Email
Subject
Comment
Verification	Please type the name of the object/variant:
File
Embed
Voice
Poll
Password	(For file deletion.)

File: 1786993964272h.png 📥︎ (102.02 KB, 2000x2000) ImgOps

Chud 09/19/26 (Sat) 17:32:46 №39364 [Quote]

Remote thin clients are the best forward for most usecases and enthusiasts and goymers can't stand it for some reason. Why are we wasting so many resources making desktops and laptops and consoles that squeeze maybe 5% of the potential of the silicon over it's lifespan? The best way to do it are giant mainframes that can take advantage of scaling, and then remoting in with devices that's only job is to stream video from and IO to the server. Their battery life could be days, they could indestructible, cheap, and immune to obsolescence.
>I don't want to be a rentcuck for compute
Self host you dumb nigger. Imagine being able to bring your desktop and everything on your home network anywhere on the planet. Imagine being able to pitch into a group of friends or family to buy a top of the line computer that you can all use at once. My dream in the future is to have 13 kids and have them all seamlessly share a dual EPYC system with like 8 GPUs that I also use for AI and whatever other homelab shit

Chud 09/19/26 (Sat) 18:12:36 №39370 [Quote]

didnt linux tech tips try dis doe
Although it could be good
Also obviously with that server, just host the entire internet for your children (manpages, package repos, muh wikipedia or whatever + muh llm since u got 999 gpus)

Chud 09/19/26 (Sat) 18:51:05 №39371 [Quote]

>>39364 (OP)
Ive been thinking of doing this instead of buying a (very expensive) laptop for my uni.
Save up for a gigaPC that i run remotely on, and even a 70B ai model. And use the old used laptop my sweet old grandmother wanted to give me to access it.
Question is, what is the appropiate hardware for this? My original plan is to just add 2x rtx 3090's to my current gaming pc i got and run with that. But it would be cool to hear some other, more scalable setups for this.

Chud 09/19/26 (Sat) 18:52:39 №39373 [Quote]

>>39371
>70b ai model
Well currently the best you could realistically run on that tier of machine would be qwen3.8-flash-next
Although for a system you could obviously have like 256gb+ of sys ram + about 32-64gb of vram (buying either rtx 3080 20gbs, or older radeon pro / nvidia tesla cards)

Chud 09/19/26 (Sat) 18:58:04 №39374 [Quote]

>>39371
Be careful with limited PCIe bandwidth and your usecase. Level1techs has some good vids about homelab AI stuff. Some things like PCIe switches are really neat

Chud 09/19/26 (Sat) 19:04:54 №39375 [Quote]

>>39371
sadly 3090s have shot up from like 500 bucks to 1000 because of muh ai jeets, yet all other cards are unaffected
avoid radeon i would say

Chud 09/19/26 (Sat) 19:05:06 №39376 [Quote]

>>39375
however you could consider it, idfk

Chud 09/19/26 (Sat) 19:15:16 №39377 [Quote]

>>39375
>>39374
>>39373
thanks for the info, il look further into it. and i meant 30b parameter models, idk why i said 70b
should i just start building the machine now? or wait for the ram prices to drop? i consulted my teacher and they said you only need the beefy hardware in 2 years. i have quite a bit of money saved up for this, but idfk if it would be retarded to spend a fortune now and barely need the compute when i could wait and spend it on better and cheaper hardware later (or get raped with whatever price influx will happen then)

Chud 09/19/26 (Sat) 19:27:40 №39378 [Quote]

File: ClipboardImage.png 📥︎ (31.14 KB, 773x134) ImgOps

File: ClipboardImage.png 📥︎ (113.04 KB, 860x438) ImgOps

>>39377
The landscape we are heading into for ai (as seen with deepseek v4.1, and qwen 4) is with muh n-gram embeddings
essentially they hash the last three tokens and then look up the hash in this table of sorts?, i dont understand it much but you can offload this table onto storage

Chud 09/19/26 (Sat) 19:31:35 №39380 [Quote]

>>39377
Currently your best bet is either
>if you have 24gb of vram
Qwen3.8-27b
>if you have over 96gb of total system memory (like 16gb of vram and 64gb+ of system memory)
qwen3.8-flash-next
>if you have like 160gb+
deepseek v4 flash-vision-exp (v4.1 is larger)
>if you have 256gb+
glm 5.3 flash
<plebbit
https://github.com/FlashML-org/FreeToken
look into dis, essentially all of these models (excl 3.8-27b) are MoE models (mixture of experts), what this does is it streams experts from the ram (in advance) to the gpu, giving similar speeds as if you were actually able to fit the whole thing on the gpu
I think if you have 32gb of vram you can run all of these (obviously with the additional system ram required

Chud 09/20/26 (Sun) 05:17:20 №39406 [Quote]

>>39370
>expecting Linus "Racemixing" Sebastian to do anything involving UNIX correctly
Yeah, it definitely could be done properly. Linux is made to be multi user after all. No hypervisor containerization bullraisin need

Chud 09/21/26 (Mon) 03:50:30 №39525 [Quote]

>>39364 (OP)
Sounds straight out of the 80s, imagine maintaining this while having a day job, kids, and pleasing your wife.

Chud 09/22/26 (Tue) 18:54:53 №39667 [Quote]

>>39525
Made me hard

Chud 09/22/26 (Tue) 20:18:03 №39696 [Quote]

nobody who types ts out is having kids keep dreaming lil bro ✌️😭

Chud 09/22/26 (Tue) 20:22:04 №39698 [Quote]

this sounds like a good idea if society was made up of ants

Chud 09/22/26 (Tue) 20:23:46 №39699 [Quote]

>>39696
i mean i will have kids and think its a great idea that will sadly never work

Chud 09/22/26 (Tue) 22:41:51 №39736 [Quote]

We need to go back to the days of computers being expensive long term investments. The push to make everything as cheap as possible raped quality, now pcs will barely last you a few years and you can't even repair them. If we normalized computers being an expensive, high quality thing that lasted awhile and could be repaired/modified then many modern issues would be solved. Sure, you'd have to put up a much larger upfront cost for one, but you wouldn't have to replace it every few years and you'd be able to do a lot more with it. Imagine having a family computer that lasted like 20 years, being constantly upgraded with newer and better parts. It would be a much better situation then our current niggerhell of paying thousands for a computer which breaks or becomes obsolete in a few years.